Frequency-Partitioned Architecture

• Frequency-Partitioned Architecture is a design approach where specific frequency bins or spectral windows serve distinct architectural roles to optimize processing.
• It enables tailored processing techniques such as separating low-frequency motion from high-frequency detail in video and allocating optimal spectral bands in OFDM and circuit-QED systems.
• This approach improves computational efficiency and performance, as seen in enhanced image quality metrics, boosted spectral efficiency, and higher quantum gate fidelities.

“Frequency-partitioned architecture” (Editor’s term) is best understood as a family of designs in which different frequency components, spectral bins, or frequency operating windows are treated as explicit architectural entities. In current arXiv usage, that family is heterogeneous rather than uniform. Some systems are literal instances: FD4MM separates low-frequency structure from high-frequency detail in video motion magnification (Wang et al., 2024); partitioned complementary sequences encode information in the gaps between OFDM frequency clusters (Sahin, 2021); frequency-resolving single-photon detectors allocate distinct subsystems to distinct spectral bins (Young et al., 2022); and microwave-driven circuit-QED architectures shift interaction selectivity into coupler-drive frequency windows (Paolo et al., 2022). Other systems are only frequency-aware, or are partitioned along channels, spatial subregions, layers, devices, or stages rather than frequency itself (Tran et al., 2024, Wu et al., 2024, Zhao et al., 2 Jan 2025).

1. Scope and taxonomy

The term has no single canonical definition across the cited literature. A useful technical distinction is between architectures that explicitly partition frequency and architectures that merely operate in the frequency domain or use some other partition axis. This distinction is explicit in several papers: CPAT is described as a “channel-partitioned attention architecture” with an “FFT-based frequency-aware interaction module,” not a strict frequency-partitioned architecture (Tran et al., 2024); PCNO is a “partitioned-in-network architecture whose coupling operator is implemented in the frequency domain,” not a model that divides computation into spectral bands (Wu et al., 2024); and the adaptive hybrid FFT partitions computation by stages, memory/dataflow roles, and parallel lanes rather than by subsets of frequency bins (Zhao et al., 2 Jan 2025).

Architectural mode	Representative paper	Partition object
Explicit low/high-frequency branching	FD4MM (Wang et al., 2024)	Low-frequency structure and high-frequency detail
Frequency-cluster signaling	Partitioned CSs (Sahin, 2021)	Nonzero OFDM clusters and zero gaps
Spectral-bin sensing	Frequency-resolving detector (Young et al., 2022)	Subsystems centered at $\omega_i$
Spectral operating-window control	Circuit-QED with variable microwaves (Paolo et al., 2022)	Qubit, coupler, and coupler-drive frequencies
Frequency-aware but not literal partitioning	CPAT (Tran et al., 2024); PCNO (Wu et al., 2024)	Channels or subregions with FFT/Fourier coupling

Taken together, these works suggest that Fourier transforms, FFT branches, or spectral convolutions are insufficient by themselves to define a frequency-partitioned architecture. The defining property is architectural specialization by spectral role: low versus high frequency, one frequency bin versus another, or one drive-frequency window versus another. Where that specialization is absent, “frequency-aware” is the more precise label.

2. Explicit low/high-frequency partitioning in learned representations

The clearest explicit example is “Frequency Decoupling for Motion Magnification via Multi-Level Isomorphic Architecture” (Wang et al., 2024). FD4MM is organized around a direct low/high-frequency split in learned feature space: $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$ where $\mathcal{W}_r$ is a dilated convolution with $r=2$ and $\vartheta$ is GELU. The paper assigns different semantic roles to the two partitions. The deep low-frequency branch carries the subtle motion field,

$\delta(x,t)=\Delta L_d(x,t),$

while the high-frequency branches preserve edges, textures, and local appearance for later reconstruction. This is a literal frequency-partitioned design because decomposition, processing, and recoupling are all organized around spectral role rather than generic multiscale hierarchy (Wang et al., 2024).

The architecture is “multi-level” because it recursively peels off high-frequency bands at shallow, middle, and deep levels, and “isomorphic” because the same decomposition pattern is reused across stages. The motion branch is processed by a Sparse Low-pass Filter with sparse attention

$\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$

then magnified through

$$L'_{d}(x,t)= L_{d}(x,t) + \mathcal{W}_{p}(\alpha \cdot \mathcal{W}_{p}(\mathcal{F}_{L}(\delta(x,t)))).$$

High-frequency bands are processed separately by Sparse High-pass Filters, and the two spectral partitions are progressively recoupled by the Sparse Frequency Mixer. The paper’s approximate image-formation expression,

$$I_{m}(x, t) \approx \sum_{i \in\{s,m,d\}} H_{i}(x,t) + \alpha \,\Delta L_{d}(x,t)\frac{\partial f(x)}{\partial x},$$

makes the division of labor explicit: detail comes from high-frequency bands, while magnified motion comes from the low-frequency branch (Wang et al., 2024).

The empirical results reinforce that the partition is architectural rather than cosmetic. Table 3 identifies three levels as the best operating point: $A_2$ achieves SSIM $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$0 and LPIPS $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$1, while the extra-band variant $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$2 degrades to SSIM $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$3 and LPIPS $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$4. Table 4 shows the cumulative effect of the partition-specific modules: the MIA baseline gives SSIM $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$5 and LPIPS $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$6, while adding $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$7, $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$8, and $$\left\{ \begin{aligned} & L(x, t) = \vartheta\cdot\mathcal{W}_{r}F(x,t), \ & H(x, t) = \vartheta\cdot(F(x,t)-\mathcal{W}_{r}F(x,t)), \end{aligned} \right.$$9 improves performance to SSIM $\mathcal{W}_r$0 and LPIPS $\mathcal{W}_r$1. The abstract further states that FD4MM “reduces FLOPs by 1.63$\mathcal{W}_r$2 and boosts inference speed by 1.68$\mathcal{W}_r$3 than the latest method” (Wang et al., 2024).

A useful contrast is CPAT. “Channel-Partitioned Windowed Attention And Frequency Learning for Single Image Super-Resolution” partitions channels into vertical, horizontal, and square-window attention paths, and adds an FFT/iFFT-based Spatial-Frequency Interaction Module. The paper is explicit that it does not perform “decomposition into separate frequency bands,” “explicit low/high-frequency branch splitting,” or “explicit partitioning of the spectrum into semantic or band-limited components” (Tran et al., 2024). That distinction marks the boundary between a literal frequency-partitioned architecture and a frequency-aware hybrid architecture.

3. Frequency-cluster signaling in OFDM

A second literal form appears in “Encoding and Decoding with Partitioned Complementary Sequences for Low-PAPR OFDM” (Sahin, 2021). Here the object being partitioned is not an internal feature tensor but the OFDM spectrum itself. A standard $\mathcal{W}_r$4-PSK complementary sequence is split into frequency clusters of nonzero subcarriers, with structured zero gaps carrying information bits. The support is generated through

$\mathcal{W}_r$5

where the phase function $\mathcal{W}_r$6 produces the standard complementary-sequence content and the integer-valued shift function

$\mathcal{W}_r$7

controls the cluster separations (Sahin, 2021).

The partitioning rule that preserves complementarity is the no-overlap condition

$\mathcal{W}_r$8

The paper shows that this rule coincides with the non-squashing partitions of a positive integer, and it introduces equivalent separation variables

$\mathcal{W}_r$9

With $r=2$0 available subcarriers and $r=2$1 nonzero tones, the zero budget is $r=2$2, and valid support patterns satisfy

$r=2$3

Architecturally, the support pattern is therefore a constrained, information-bearing frequency partition rather than an arbitrary sparse allocation (Sahin, 2021).

The main design objective is to increase spectral efficiency without losing the low-PAPR property of complementary-sequence OFDM. Because the construction remains within the CS family, the paper states that the maximum PAPR is still $r=2$4 dB. The total number of signaling options becomes

$r=2$5

where $r=2$6 counts admissible support patterns and $r=2$7 is the number of standard $r=2$8-PSK complementary sequences. The paper’s numerical example is representative: for $r=2$9, standard CS with $\vartheta$0, $\vartheta$1, yields about $\vartheta$2 bits, whereas partitioned CS with $\vartheta$3, $\vartheta$4, yields about $\vartheta$5 bits (Sahin, 2021).

This gain is not free. The decoder becomes support-aware and recursive, and minimum distance depends on both the content code and the support set. The paper therefore introduces a minimum-distance-constrained subset indexed by $\vartheta$6, with

$\vartheta$7

and the bound

$\vartheta$8

The reported tradeoff is that unrestricted partitioning raises SE but incurs limited SNR loss, while constrained partitioning can perform similarly to standard CSs in average BLER while retaining higher SE (Sahin, 2021).

4. Spectral operating windows in superconducting quantum architectures

In superconducting circuit-QED, frequency partitioning takes a different form: not subbands of a signal, but controlled allocation of qubit, coupler, and drive frequencies. “Extensible circuit-QED architecture via amplitude- and frequency-variable microwaves” combines fixed-frequency qubits with microwave-driven couplers and explicitly argues that interaction selectivity can be shifted from qubit tunability into “qubit frequencies, coupler frequencies, and coupler-drive operating windows” (Paolo et al., 2022). The logical ZZ interaction is defined from quasienergies as

$\vartheta$9

and gate design is framed as moving through amplitude/frequency space so that $\delta(x,t)=\Delta L_d(x,t),$0 is small at idle and strong during the gate.

The paper’s central systems idea is a two-step frequency allocation: first choose qubit and coupler frequencies to minimize undesired static ZZ couplings; then choose coupler-drive frequencies to maximize desired driven ZZ while minimizing spectator activation. The authors state that they “take advantage of the approximate decoupling between static and driven interactions in the proposed architecture, and optimize the parameters using a two-step process” (Paolo et al., 2022). That is a literal spectral partition of architectural roles: one frequency layer for idle operation, another for active control. The abstract reports estimated average gate fidelities beyond $\delta(x,t)=\Delta L_d(x,t),$1 with gate times in the range $\delta(x,t)=\Delta L_d(x,t),$2-$\delta(x,t)=\Delta L_d(x,t),$3, and the body describes driven ZZ operating over a large drive-frequency bandwidth (Paolo et al., 2022).

“Tunable Coupling Architecture for Fixed-frequency Transmons” provides a closely related, though differently implemented, lesson (2101.07746). It does not use the phrase frequency-partitioned architecture explicitly, but it studies how fixed-frequency qubits can remain in favorable frequency classes while a tunable coupler supplies on/off interaction control. The architectural move is the bus-below-qubits (BBQ) regime, in which the coupler frequency is below the qubit frequencies at idle. The paper reports an off-state conditional interaction

$\delta(x,t)=\Delta L_d(x,t),$4

a best CZ gate time of $\delta(x,t)=\Delta L_d(x,t),$5, and interleaved-RB error per gate $\delta(x,t)=\Delta L_d(x,t),$6, corresponding to $\delta(x,t)=\Delta L_d(x,t),$7 fidelity (2101.07746).

Taken together, these two papers suggest a quantum variant of frequency partitioning in which fixed-frequency qubits define a static spectral backbone, while coupler placement and coupler-drive windows define localized interaction channels. The resulting architecture is neither purely frequency-static nor purely flux-tunable; it partitions spectral responsibilities across qubit frequencies, coupler frequencies, and control-drive frequencies (Paolo et al., 2022, 2101.07746).

5. Spectral-bin partitioning in frequency-resolving photodetection

“Nanoscale Architecture for Frequency-Resolving Single-Photon Detectors” is a direct example of a frequency-partitioned sensing system (Young et al., 2022). The detector consists of $\delta(x,t)=\Delta L_d(x,t),$8 subsystems, each centered at a subsystem-specific frequency $\delta(x,t)=\Delta L_d(x,t),$9, and each with its own readout channel. The paper states that the detector is “a single detector composed of multiple spectrally distinct sub-detectors,” and that the output channel that fires identifies the photon’s frequency bin. This is not a dispersive spectrometer that routes wavelengths to separate detectors; the whole subwavelength device interacts collectively with a common optical mode (Young et al., 2022).

The collective light-matter coupling is expressed through

$\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$0

and the long-time click probability for output channel $\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$1 is denoted $\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$2. Frequency estimation is then derived from the output distribution through

$\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$3

with $\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$4 used as the frequency-resolution measure (Young et al., 2022). Architecturally, the subsystems are frequency bins, the click distribution is a spectral partition map, and the cooperative interaction means the bins are not independent.

The performance regime reported in the abstract is unusually strong: frequency resolution over a $\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$5 eV bandwidth in the visible range, near perfect detection efficiency, jitter of a few hundred femtoseconds, and frequency resolution of tens of meV (Young et al., 2022). The detailed tradeoff curve is also explicit: one operating point gives $\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$6 fs jitter with $\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$7 meV resolution, while a more spectrally optimized regime gives $\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$8 meV resolution with only $\mathbf{SA}^{(h)}_{L}= \operatorname{ReLU}\!\left( \frac{\mathbf{Q}_{L} \mathbf{K}^{\mathrm{T}}{\tau}\right),$9 fs jitter. The paper further shows that adding unmonitored end subsystems suppresses unwanted edge absorption, which illustrates a distinctive consequence of cooperative frequency partitioning: not every spectral partition must correspond to an output channel in order to shape the global response (Young et al., 2022).

The proposed physical realization uses carbon nanotubes functionalized with quantum dots. Different CNT/QD groups implement different transition energies and therefore different spectral bins, while CNT conductivity changes provide the readout signal. This realization preserves the central architectural idea: frequency is resolved by internal spectrally specialized subsystems rather than by external wavelength dispersion (Young et al., 2022).

6. Adjacent concepts, boundary cases, and common confusions

A recurrent source of confusion is the assumption that any architecture with Fourier processing, FFT blocks, or “partitioned” components is therefore frequency-partitioned. The cited literature repeatedly rejects that equivalence. CPAT uses FFT and iFFT inside the Spatial-Frequency Interaction Module, but the paper explicitly states that the model does not specify “decomposition into separate frequency bands,” “explicit low/high-frequency branch splitting,” or “learnable frequency partition indices”; its explicit partition is over channels, not frequencies (Tran et al., 2024). PCNO performs joint convolution in the Fourier layer and requires grid alignment so regions share a common Fourier basis, yet it is described as “partitioned-in-space / partitioned-in-network” rather than frequency-partitioned because the explicit partitions are physical subregions, not spectral bands (Wu et al., 2024).

The same boundary applies to systems partitioned by runtime orchestration or hardware resource. “Joint Partitioning and Placement of Foundation Models for Real-Time Edge AI” partitions contiguous layer/block segments $$L'_{d}(x,t)= L_{d}(x,t) + \mathcal{W}_{p}(\alpha \cdot \mathcal{W}_{p}(\mathcal{F}_{L}(\delta(x,t)))).$$0 and jointly optimizes segmentation and placement, but it explicitly does “not” partition by frequency bands, spectral components, or Fourier components (Djuhera et al., 30 Nov 2025). “Bandwidth Allocation with Device Partitioning for Federated Learning over Industrial IoT networks” likewise partitions devices into ordered subsets that receive exclusive access to the full bandwidth $$L'_{d}(x,t)= L_{d}(x,t) + \mathcal{W}_{p}(\alpha \cdot \mathcal{W}_{p}(\mathcal{F}_{L}(\delta(x,t)))).$$1 sequentially; the paper stresses that this is “not static frequency partitioning” but “time-sequential bandwidth exclusivity” (Kim et al., 29 May 2026). “Adaptive Hybrid FFT” partitions computation by stages, banks, and dataflow mode rather than by disjoint groups of output frequencies (Zhao et al., 2 Jan 2025).

These distinctions matter because they separate three concepts that are often conflated. A frequency-partitioned architecture assigns different spectral entities different architectural roles. A frequency-aware architecture uses FFTs, Fourier layers, or frequency-domain processing but may not partition the spectrum into roles. A non-frequency partitioned architecture may be partitioned just as deeply—by channels, space, devices, layers, or hardware stages—without spectral specialization being the organizing principle. The cited corpus suggests that keeping these categories separate is necessary for technical precision (Tran et al., 2024, Wu et al., 2024, Djuhera et al., 30 Nov 2025, Kim et al., 29 May 2026, Zhao et al., 2 Jan 2025).